How it works

Keep financial truth deterministic; use AI only where the data is messy, and only from read-only evidence.

Pipeline

Sources

Payments, refunds, gateway settlements, bank statement, ledger export. Insert-only, or attached read-only from the customer's export.

Stage 1 · Payment ↔ settlement item

Exact id link. Fee and 18% GST on the fee re-computed in integer paise; any difference is an exception.

Stage 2a · Exact UTR

Settlement UTR found in the bank narration after normalising case, spaces and hyphens.

Stage 2c · Fuzzy UTR

Edit distance on the UTR: ≤ 1 with a clear margin auto-matches; 2 or truncated is only suggested. Exact amount and date are required.

Stage 2b · Batch decomposition

A unique run of consecutive settlements, else a proven-unique subset (meet-in-the-middle / budgeted branch-and-bound). Otherwise: ranked candidate sets for a human.

Stage 3 · Bank ↔ ledger

Receipts by reference, then amount + date + party. Duplicates, mismatches and missing entries; a period tie-out must explain every rupee.

Stage 4 · Timing

Per-method T+n windows on Indian business days (Sundays, 2nd and 4th Saturdays, holidays). Late is always reported.

Output

Matches, exceptions with machine-readable rule traces, exposure per class, the tie-out, and an append-only audit trail.

Trust boundary

Deterministic engine

Decides every match, exception and exposure class
All money is integer paise; no floats
Auto-matches a batch only with a uniqueness proof (or a stated contiguity assumption)
Financial records are insert-only and can be mounted read-only

Constrained AI layer

Reads a read-only evidence bundle; never writes financial data
May suggest UTR tokens only if they literally appear in the narration
Explanations citing evidence outside the bundle are rejected
Draft journals must balance to the engine's amount; status pending_approval
Any failure falls back to a deterministic template, labelled as such

Human reviewer

Approves suggestions and picks batch candidate sets
Two steps: preview + 5-minute token, then token + name
Every action lands in the append-only audit trail
Approving a set re-checks the leftover settlements

Run the full agent with approvals (MCP)

This site is a read-only snapshot. The agent runs locally over stdio.

npx -y closeloop-mcp          # MCP server; seeds demo data on first run
npx -y closeloop-mcp check    # one reconciliation, summary to stderr

Claude Desktop config:

{ "mcpServers": { "closeloop": { "command": "npx", "args": ["-y", "closeloop-mcp"] } } }

Tools:

run_reconciliationlist_exceptionsget_evidenceexplain_exceptiondecompose_bank_creditdraft_journal_entryapprove_suggestionescalateget_audit_trail

Plus the month_end_review prompt and the closeloop://summary/latest resource. Every amount arrives pre-formatted with totals, so the model never does arithmetic.

Demo

Screenshot or GIF slot: add apps/dashboard/public/demo.gif (or .png / .webp) and rebuild.